NVIDIA's Nemotron 3.5 Transcribe tops open ASR leaderboard at 5.33% English WER
Victoria Shinder1 min readNVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model on Hugging Face and ModelScope that combines transcription, speaker attribution, translation and audio question answering in one checkpoint. According to the announcement, it transcribes 40 languages, translates 31 of them to or from English, attributes speakers in 13, handles recordings up to an hour long and labels up to eight speakers. Output is steered with natural-language prompts, for example a list of expected medical terms to use only when the audio supports them.
On the Hugging Face Open ASR Leaderboard it is first among open-weight models for English at 5.33% average word error rate, with API models excluded. NVIDIA also cites 4.62% mean WER across seven languages and 10.98% on long-form audio, both the lowest in the model card's comparison, and the lowest error rate on David AI's 21-language diarization leaderboard. Word-level timestamps are supported; audio question answering is trained in English only. Quantized versions are announced as coming.
Alongside the model, NVIDIA published Granary v2, a multilingual dataset and curation pipeline, plus NeMo fine-tuning recipes. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt. The post itself advises spot-checking names, speaker changes and answers against the recording.
What it means
The practical change is consolidation: diarization, translation and transcription usually mean two or three models chained together, each with its own failure modes. One promptable checkpoint simplifies that pipeline. The leaderboard positions are NVIDIA's reading of public boards, and the medical result compares fine-tuning against prompting on the same model, which says more about the value of domain data than about rivals.